Crafting your own RAG system: Leveraging 30+ LLMs for enhanced performance
summary
Prototype shows a Java RAG pipeline that indexes files and GitHub repos, experiments with 30+ LLMs (local and cloud) and compares embeddings, chunking and reranking to reduce hallucinations; it reports surprises like small embeddings outperforming the default and limits (PGVector caps at 2,000 dims) and demonstrates LangChain for Java integration, a local AMA/GPT4All setup, and binary embeddings to cut storage dramatically while trading tiny accuracy loss.
AI-written summary · about
chapters (12)
- 0:00Introduction and motivation
- 3:36Debugging RAG prototype journey
- 5:23RAG flow and architecture overview
- 9:39Advanced RAG components
- 13:27Prototype development with 30+ LLMs
- 19:55Demo of the RAG prototype
- 24:09Vector databases and embeddings
- 33:03Embedding dimensions and storage
- 37:45Lessons learned and evaluation
- 40:15Devox Genie plugin introduction
- 46:46Future possibilities and conclusion
- 47:08Q&A
related talks
48mCrafting your own RAG system: Leveraging 30+ LLMs for enhanced performance
56mRAG: from dumb implementation to serious results
42mTo RAG or not to RAG
28mNavigating Challenges and Enhancing Performance of LLM based Applications
27mOptimizing RAG Systems in Practice: Challenges and Proven Solutions
01h 33mRAG How To Talk To Your Data
02h 47mFrom naive to advanced RAG: the complete guide
19mRAG in 2025: State of the Art and the Road Forward